Dual core speed increasement?

Hi there all!

It has been a while since anyone published something in here. (We really should use this section more often...)
I was curious what experience others have with Dual-core speed increasements. Below you can see the increasement in the newest test-version of Brutus:

Code:


Single core
 7  -4  229  5806826 c2c4
 depth=7
 moves searched=15
 Nodes=12303458
 average speed= 2458.kNPS  total time elapsed= 5004
 bestmove=c2c4

Dual core
 depth=7
 moves searched=43
 Nodes=18916618
 average speed= 3776.kNPS  total time elapsed= 5009
 bestmove=c2c4


The increasement is 3776/2458 * 100% = 153%

Even if I start 2 seperate games and count the nodes together, I end up getting 150% of the single-core speed at Max. What is the experience of others regarding this?
I am using an AMD-dual core 4600. Is pentium better than this (and why?) Feel free to test Brutus on YOUR dualcore pentium, I have compiled a Win32 version, just start the engine and type "mt", it will give output like above. You can download it from:
http://home.casema.nl/vermeire/Brutus7_test037.zip
(By the way, it is a fully operational winboard version, very fast and yet very weak because it lacks a descent evaluation).

Any input is welcome!
Stephan

Hi Stephan,

I just ran your Brutus7_test037 "mt" gives the following aoutput on an core2duo E6300 @2.1Ghz:

Code:

SINGLE CORE
 1  -188  0  13 f6e4
 1  -183  0  23 e7c5
 1  -80  0  52 b7b6
 1  13  0  62 g7g6
 1  16  0  135 f8e8
 2  10  0  350 f8e8
 3  10  0  3216 f8e8
 4  7  1  16278 f8e8
 4  10  2  40396 a5a4
 5  15  4  87121 a5a4
 5  16  7  161569 f6d7
 6  17  12  306838 f6d7
 7  23  56  1414084 f6d7
 8  18  440  11896894 f6d7
 depth=8
 moves searched=2
 Nodes=13441660
 average speed= 2685.kNPS  total time elapsed= 5005
 bestmove=f6d7

DUAL CORE
 1  -188  0  13 f6e4
 1  -80  0  46 b7b6
 1  12  0  65 h7h6
 1  13  0  76 g7g6
 1  16  0  147 f8e8
 2  10  0  381 f8e8
 3  10  0  3971 f8e8
 4  7  1  18969 f8e8
 4  10  2  47959 a5a4
 5  15  3  95884 a5a4
 5  16  6  173708 f6d7
 6  17  10  348173 f6d7
 7  23  36  1455294 f6d7
 8  18  204  9525296 f6d7
 depth=8
 moves searched=22
 Nodes=23181436
 average speed= 4632.kNPS  total time elapsed= 5004
 bestmove=f6d7

Seems to scale a bit better.

All the best
Michael Schmidt

Hi Michael,

Thanx for benching the test-version on your machine! Indeed it scales better on your machine: Duo = 172.5% of the single thread. I did some more experimenting, it turns out to have something to do with Intel-compiler optimalisations that perform not as well in a multi-thread. I also found another compile-option that improves performance:

B7038PGO+parallel: 150%
B7038PGO+parallel: 161%

If I extend the evaluation-function, it does less alphaBeta loops and the scaling improves even more:

B7038PGO+parallel FullEval: 172%

But then again, that is where you started! I'm intending to investigate multiprocessing now instead of multithreading. THAT will sure fix these problems and give a whole load of others!!!

Stephan

Just run br7 on my quad:

bestmove: f6d7

single: 3126
dual: 4982

ratio: 1.5937

best,

Thanx for benching the test-version Harun!

Currently I am working on a multi-process version. As soon as I have some results, I will post them here for those who are interested. One way or another, the scaling MUST do a little better.

Stephan